DeepSeek's current model is DeepSeek-V4.1-Flash, released on September 10, 2026 and served on the API under the name deepseek-flash. It is the most recent DeepSeek release. The larger DeepSeek-V4-Pro, which reached general availability on August 13, 2026, stays in service until a V4.1-Pro launches. DeepSeek has not given a date for that. This page lists each DeepSeek release from V2 onward with dates taken from DeepSeek's own API changelog.
DeepSeek Model Timeline
| Model | Released | Context window | Notable change | Status |
|---|---|---|---|---|
| DeepSeek-V2 | May 6, 2024 | 128K | 236B mixture-of-experts model, 21B active per token | Discontinued |
| DeepSeek-V2.5 | September 5, 2024 | Not disclosed | Merged the chat and coder lines into one model | Superseded |
| DeepSeek-V3 | December 26, 2024 | 128K on the model card | 671B parameters, 37B active. Open weights. Updated on March 24, 2025 | Weights available, replaced on the API |
| DeepSeek-R1 | January 20, 2025 | Not stated in the launch post | Open reasoning model under the MIT licence. Updated on May 28, 2025 | Weights available, replaced on the API |
| DeepSeek-V3.1 | August 21, 2025 | 128K | One model with thinking and non-thinking modes | Superseded |
| DeepSeek-V3.1-Terminus | September 22, 2025 | Not disclosed | Fixes for language consistency and agent behaviour | Superseded |
| DeepSeek-V3.2 | December 1, 2025 | Not disclosed | Followed the V3.2-Exp build of September 29, 2025. Both API names moved to it | Superseded |
| DeepSeek-V4 preview (Flash and Pro) | April 24, 2026 | 1M | Flash: 284B parameters, 13B active. Pro: 1.6T parameters, 49B active. 1M context became the default on all DeepSeek services | Replaced by official releases |
| DeepSeek-V4-Flash | July 31, 2026 | 1M | Official release, retrained after the preview with the same architecture | Retired. Requests now served by V4.1-Flash |
| DeepSeek-V4-Pro | August 13, 2026 | 1M | General availability on app, web, and API. Three thinking effort levels | Active |
| DeepSeek-V4.1-Flash | September 10, 2026 | 1M, with 384K maximum output | 552B parameters, 8B active on input and 16B on output. Native image input | Current |
The legacy API names deepseek-chat and deepseek-reasoner were scheduled for retirement on July 24, 2026, three months after the V4 preview. Before V2, DeepSeek released DeepSeek-Coder on November 2, 2023 and DeepSeek-LLM on November 29, 2023.
API Pricing Over Time
| Model and date | Input, cache miss | Output | Source |
|---|---|---|---|
| V3, list price from February 8, 2025 | $0.27 | $1.10 | DeepSeek launch post |
| R1, January 2025 | $0.55 | $2.19 | DeepSeek launch post |
| V4-Flash preview, April 2026 | $0.14 | $0.28 | CloudZero pricing guide |
| V4-Pro preview, April 2026 | $1.74 list, $0.435 under a 75 percent launch discount | $3.48 list, $0.87 discounted | CloudZero pricing guide |
| V4-Pro, from August 16, 2026 | $0.66 off-peak, $1.32 peak | $1.98 off-peak, $3.96 peak | DeepSeek pricing page |
| V4.1-Flash, from September 10, 2026 | $0.15 off-peak, $0.30 peak | $0.60 off-peak, $1.20 peak | DeepSeek pricing page |
Prices are per million tokens in US dollars. Since August 16, 2026 DeepSeek bills by time of day: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and off-peak rates are half the peak rate. Cached input on V4.1-Flash costs $0.003 off-peak and $0.006 at peak.
Open Weights
DeepSeek has published weights for the main models in this table, including both V4 models. R1 and V3.1 onward use the MIT licence, and V4.1-Flash is on Hugging Face under MIT with its technical report. V3 shipped with MIT-licensed code and a separate model licence that allows commercial use.
What Is Next
V4.1-Pro. DeepSeek's September 10 announcement calls V4.1-Flash the smallest model in a new architecture family and refers to a V4.1-Pro launch without dating it. The same post said V4-Pro requests would be routed to V4.1-Flash from September 14. DeepSeek's changelog and pricing page now show V4-Pro still served and billed on its own, so that plan was not carried out as first written.
Anything beyond V4.1. We found no DeepSeek statement about a V5 or a new R-series model. The R line ended in practice with V3.1, which folded reasoning into the main model.
Brand Visibility Implications
- DeepSeek does not publish knowledge cutoffs for these models. The V4.1-Flash model card gives none. Brand recall has to be tested, not looked up.
- Old names now point at new models. A tool that still calls deepseek-v4-flash gets V4.1-Flash. Answers about your brand can change with no change on the tool's side. Record the model version with every tracked answer.
- Open weights outlive the API. V3 and R1 remain downloadable under permissive licences and can still be run in self-hosted products. Those copies keep their original training data indefinitely.
- Low prices spread the model widely. At $0.15 input off-peak, V4.1-Flash is cheap enough to sit behind developer tools and internal assistants where no retrieval is attached. See DeepSeek citation patterns and how training cutoffs change brand recall.
Methodology
This page is built from vendor documentation and dated reporting, not from Presenc AI measurements. Primary sources are used wherever one exists: OpenAI's model pages and deprecations page, Google's Gemini API changelog and deprecations page, xAI's release notes and model list, DeepSeek's API changelog and pricing page, and MiniMax's pricing page and Hugging Face model cards. Older release dates that vendors no longer document come from the relevant Wikipedia articles. Where a figure comes only from a secondary tracker or a single outlet, the page names it. "Not disclosed" means the vendor material we read does not state the figure. Status as of October 1, 2026.
How Presenc AI Helps
Presenc AI tracks how DeepSeek and other models describe your brand and records which model version produced each answer. When a vendor reroutes an old model name to a new model, you can see whether the answers moved.